Let me summarize the entire day into one slide.

When Claude is generating code, analyzing your code, or producing any kind of output, it's working with two things: its own internal knowledge from training, and a sequence of tokens that make up the context. It's predicting the next word - or more precisely, the next token - based on what it's seen.

Your task is to manage that context effectively. It needs to be four things: Complete, Relevant, Right, and Tight.

Complete means Claude needs enough context to know what you're asking it to do, and to understand the full cause and consequence of whatever it tries to do.

Relevant means every token in that context contributes to the probability distribution of the next output. If there's off-topic stuff in the context, it decreases the probability of a sensible, useful output and increases the probability of something completely irrelevant.

Right means the context needs to be accurate. If Claude makes a mistake, that mistake becomes part of the context, which lowers the probability of a correct answer and increases the probability of more mistakes. This is important: if Claude makes a mistake based on something you said, there's no point trying to correct it. You're better off rewinding back to before the mistake and restarting with better information. That's weird, because it's not how human beings work, but it's how Claude works.

Finally, Tight. When you first start a Claude Code session, there's almost nothing in context, so it's lightning fast and inexpensive. Cost is related to context size, so keeping it tight saves time and money. Of course, there's a trade-off: being as complete as possible means you're not being as tight as possible.

This changes how we think about programming. First, abundance. We're so used to code being expensive that we don't want to throw it away. But with Claude, it's different. If it generates rubbish, you can just run it again ten times and take the best one. It doesn't have an ego problem like a human would.

Second, autonomy. The human being is the bottleneck to abundance. So how do you get out of the way? You need test suites so Claude can write code, run tests, see results, and create pull requests without you being involved. You need hooks to make sure it's coding to the right standards and triggers to block it from doing stupid things. You're providing oversight while Claude works autonomously within those parameters.